Inpainting micro-CT images of fibrous materials using deep learning

Abstract

Micro-computed tomography (CT) is an irreplaceable tool to characterize the three-dimensional microstructure of fiber-reinforced composites and other fibrous materials. However, this technique has limitations due to the small specimen size, image artifacts, and shaded regions of the image. With the development of the inpainting and deep learning techniques, especially with generative adversarial networks, it has become possible to effectively reconstruct missing or damaged parts of images, or to create a new (part of) image, which would twin the studied material. The literature, however, lacks methods that work with three-dimensional (3D) micro-CT images. In this work, we propose the deep learning-based methodology for the inpainting of 3D images of fibrous materials using a limited number of micro-CT scans. Three different architectures of 3D encoder-decoder generative adversarial network were designed and analyzed for 3D micro-CT images of random fiber composites to inpaint the artificially missing data. The developed algorithms are validated using images of compression molded short glass fiber reinforced thermoplastic composite. Inpainting results show good performance of the methods using the both image-related and physical quality metrics.